tricky query with patterns of consecutive rows

Question

Post reply

tricky query with patterns of consecutive rows

rotcha99

Mr or Mrs. 500

Points: 594

January 17, 2022 at 12:29 pm

Go to Answer

#3973846

Hi!

I'm having struggle with selecting rows, that match "dynamic" rules as described:

if there are 10 or more rows containing a value greater than 1000 this is my START signal
if there are 10 or more rows containing a value smaller than 1000 this is my STOP signal
i want everything between a START and a STOP block

I think it has to do with "gaps and islands" and probably something with lead/lag to build group ids. But I don't know how to build the "counter" so I know if it's 10 or more records that match the pattern.

Please see this simplified demo table (id is an INT, in real-world it's a timestamp of an event)

drop table if exists patternDemo;

create table patternDemo
(
    id int,
    value int,
    include char(3) -- select record in final query?
)


-- ignore everything before START
insert into patternDemo values (1, 100, 'no');
insert into patternDemo values (2, 100, 'no');
insert into patternDemo values (3, 100, 'no');
insert into patternDemo values (4, 200, 'no');
insert into patternDemo values (5, 100, 'no');
insert into patternDemo values (6, 300, 'no');
insert into patternDemo values (7, 500, 'no');
-- START: block having >= 10 consecutive valus above 1000
insert into patternDemo values (8, 1500, 'yes');
insert into patternDemo values (9, 1500, 'yes');
insert into patternDemo values (10, 1300, 'yes');
insert into patternDemo values (11, 1600, 'yes');
insert into patternDemo values (12, 1500, 'yes');
insert into patternDemo values (13, 1700, 'yes');
insert into patternDemo values (14, 1800, 'yes');
insert into patternDemo values (15, 1800, 'yes');
insert into patternDemo values (16, 1800, 'yes');
insert into patternDemo values (17, 1800, 'yes');
insert into patternDemo values (18, 1600, 'yes');
insert into patternDemo values (19, 1500, 'yes');
-- after a START block having at least 10 rows with a value over 1000
-- there maybe lower values, which i want to include, since we're still
-- before a STOP pattern
insert into patternDemo values (20, 600, 'yes');
insert into patternDemo values (21, 400, 'yes');
insert into patternDemo values (22, 1500, 'yes');
insert into patternDemo values (23, 1800, 'yes');
insert into patternDemo values (24, 1800, 'yes');
insert into patternDemo values (25, 1800, 'yes');
-- STOP pattern: another block of at least 10 rows with a value lower than 1000
insert into patternDemo values (26, 300, 'no');
insert into patternDemo values (27, 300, 'no');
insert into patternDemo values (28, 100, 'no');
insert into patternDemo values (29, 100, 'no');
insert into patternDemo values (30, 100, 'no');
insert into patternDemo values (31, 100, 'no');
insert into patternDemo values (32, 100, 'no');
insert into patternDemo values (33, 100, 'no');
insert into patternDemo values (34, 100, 'no');
insert into patternDemo values (35, 100, 'no');
insert into patternDemo values (36, 100, 'no');
insert into patternDemo values (37, 100, 'no');
insert into patternDemo values (38, 100, 'no');
insert into patternDemo values (39, 100, 'no');
insert into patternDemo values (40, 100, 'no');

Ken McKelvey

SSCoach

Points: 18910

January 17, 2022 at 6:25 pm

Answer

#3973966

This should cope with more edge conditions than my original query:

WITH Rules
AS
(
   SELECT id, [value]
      ,CASE
         WHEN 10 = SUM(CASE WHEN [value] >= 1000 THEN 1 ELSE 0 END)
                  OVER (ORDER BY id ROWS BETWEEN CURRENT ROW AND 9 FOLLOWING)
         THEN 1
         ELSE 0
      END AS StartSignal
      ,CASE
         WHEN 10 = SUM(CASE WHEN [value] < 1000 THEN 1 ELSE 0 END)
                  OVER (ORDER BY id ROWS BETWEEN CURRENT ROW AND 9 FOLLOWING)
         THEN 1
         ELSE 0
      END AS StopSignal
   FROM patternDemo
)
,StartStop
AS
(
   SELECT id, [value]
      ,CASE
         WHEN 1 = LAG(StartSignal) OVER (ORDER BY id)
         THEN 0
         ELSE StartSignal
      END AS StartSignal
      ,CASE
         WHEN 1 = LAG(StopSignal) OVER (ORDER BY id)
         THEN 0
         ELSE StopSignal
      END AS StopSignal
      ,MAX(StartSignal) OVER (ORDER BY id)
         -MAX(StopSignal) OVER (ORDER BY id) AS StartStopDiff
   FROM Rules
)
,StartWithStart
AS
(
   SELECT id, [value], StartSignal
      ,CASE
         WHEN StartStopDiff = -1
         THEN 0
         ELSE StopSignal
      END AS StopSignal
   FROM StartStop
)
,Includes
AS
(
   SELECT id, [value]
      ,SUM(StartSignal) OVER (ORDER BY id)
         - SUM(StopSignal) OVER (ORDER BY id) AS IncludeSignal
   FROM StartWithStart
)
SELECT id, [value], IncludeSignal
FROM Includes
WHERE IncludeSignal >= 1;

This reply was modified 3 years, 4 months ago by Ken McKelvey.
This reply was modified 3 years, 4 months ago by Ken McKelvey.

Viewing 6 posts - 1 through 6 (of 6 total)

You must be logged in to reply to this topic. Login to reply

DesNorton SSC-Insane Points: 24511 More actions · Answer 1

Well this is definitely a head scratcher. The following code is not pretty, but it seems to get the expected results

WITH cteBase AS (
  SELECT  *
        , rn           = ROW_NUMBER() OVER ( ORDER BY pd.id )
        , IsStartBlock = SUM( CASE WHEN pd.VALUE > 1000 THEN 1 ELSE 0 END )
                         OVER ( ORDER BY pd.id ROWS BETWEEN 9 PRECEDING AND CURRENT ROW )
        , IsEndBlock   = SUM( CASE WHEN pd.VALUE < 1000 THEN 1 ELSE 0 END )
                         OVER ( ORDER BY pd.id ROWS BETWEEN CURRENT ROW AND 9 FOLLOWING )
  FROM    #patternDemo AS pd
)
, cteStart AS (
  SELECT  sb.*
        , grp = sb.rn - ROW_NUMBER() OVER ( ORDER BY sb.rn )
  FROM    cteBase AS sb
  WHERE   sb.IsStartBlock IN ( 1, 10 )
)
, cteStartBlock AS (
  SELECT      sb.StartBlockStart
            , StartBlockEnd = MAX( st.id )
            , rn = ROW_NUMBER() OVER (ORDER BY sb.StartBlockStart)
  FROM        cteStart  AS st
  OUTER APPLY ( SELECT  StartBlockStart = MAX( sbs.id )
                FROM    cteStart AS sbs
                WHERE   sbs.IsStartBlock = 1
                  AND   sbs.id < st.id
              ) AS sb
  WHERE       st.IsStartBlock = 10
  GROUP BY    st.grp, sb.StartBlockStart
)
, cteEnd as (
  SELECT  eb.*
        , grp = eb.rn - ROW_NUMBER() OVER ( ORDER BY eb.rn )
  FROM    cteBase AS eb
  WHERE   eb.IsEndBlock IN ( 1, 10 )
)
, cteEndBlock AS (
  SELECT      EndBlockStart = MIN( nd.id )
            , eb.EndBlockEnd
            , rn = ROW_NUMBER() OVER (ORDER BY eb.EndBlockEnd)
  FROM        cteEnd  AS nd
  OUTER APPLY ( SELECT  EndBlockEnd = MIN( ebs.id )
                FROM    cteEnd AS ebs
                WHERE   ebs.IsEndBlock = 1
                  AND   ebs.id > nd.id
              ) AS eb
  WHERE       nd.IsEndBlock = 10
  GROUP BY    nd.grp, eb.EndBlockEnd
)
, cteBlocks AS (
  SELECT    BlockNo = scte.rn
          , scte.StartBlockStart
          , scte.StartBlockEnd
          , ecte.EndBlockStart
          , ecte.EndBlockEnd
  FROM      cteStartBlock AS scte
  LEFT JOIN cteEndBlock   AS ecte
         ON ecte.rn = scte.rn
)
SELECT      bk.BlockNo
          , src.*
FROM        #patternDemo AS src
INNER JOIN  cteBlocks    AS bk
        ON  src.id > bk.StartBlockEnd
       AND  src.id < bk.EndBlockStart
ORDER BY src.id;

How to post data/code on a forum to get the best help.

Ken McKelvey SSCoach Points: 18910 More actions · Answer 2

WITH Rules
AS
(

SELECT id, [value]
,CASE
WHEN 10 = SUM(CASE WHEN [value] >= 1000 THEN 1 ELSE 0 END)
OVER (ORDER BY id ROWS BETWEEN CURRENT ROW AND 9 FOLLOWING)
THEN 1
ELSE 0
END AS StartSignal
,CASE
WHEN 10 = SUM(CASE WHEN [value] < 1000 THEN 1 ELSE 0 END)
OVER (ORDER BY id ROWS BETWEEN CURRENT ROW AND 9 FOLLOWING)
THEN 1
ELSE 0
END AS StopSignal
FROM patternDemo
)
,StartStop
AS
(
SELECT id, [value]
,CASE
WHEN 1 = LAG(StartSignal) OVER (ORDER BY id)
THEN 0
ELSE StartSignal
END AS StartSignal
,CASE
WHEN 1 = LAG(StopSignal) OVER (ORDER BY id)
THEN 0
ELSE StopSignal
END AS StopSignal
FROM Rules
)
,StartWithStart
AS
(
SELECT id, [value], StartSignal
,CASE
WHEN LAG(StopSignal) OVER (ORDER BY id) IS NULL
THEN 0
ELSE StopSignal
END AS StopSignal
FROM StartStop
)
,Includes
AS
(
SELECT id, [value]
,SUM(StartSignal) OVER (ORDER BY id)
- SUM(StopSignal) OVER (ORDER BY id) AS IncludeSignal
FROM StartWithStart
)
SELECT id, [value]
FROM Includes
WHERE IncludeSignal = 1;

Steve Collins SSCrazy Points: 2214 More actions · Answer 3

It seems safer to avoid LEAD and LAG because start/stop conditions might repeat within ranges of groups

with 
gap_cte as (
    select *, iif(v.sgn<>lag(v.sgn) over (order by id), 1, 0) gap
    from #patternDemo pd
         cross apply (values (sign([value]-1000))) v(sgn)),
grp_cte as (
    select *, sum(gap) over (order by id) grp
    from gap_cte),
rn_cte as (
    select *, 
           row_number() over (partition by grp order by id desc) rn,
           row_number() over (partition by grp order by id) rn_max
    from grp_cte),
start_stop_cte as (
    select top(1) with ties *
    from rn_cte
    where rn_max>=10
    order by rn)
select g.*
from start_stop_cte s
     outer apply (select top(1) grp
                  from start_stop_cte ss
                  where ss.id<s.id
                        and ss.sgn=1
                  order by id desc) starts(start_grp)
     left join grp_cte g on g.grp<s.grp 
                            and g.grp>=starts.start_grp
where s.sgn=-1
order by rn;

Aus dem Paradies, das Cantor uns geschaffen, soll uns niemand vertreiben können

Ken McKelvey SSCoach Points: 18910 More actions · Answer 4

On looking at this again this may be a safer approach:

WITH Rules
AS
(
   SELECT id, [value]
      ,CASE
         WHEN 10 = SUM(CASE WHEN [value] >= 1000 THEN 1 ELSE 0 END)
                  OVER (ORDER BY id ROWS BETWEEN CURRENT ROW AND 9 FOLLOWING)
         THEN 1
         ELSE 0
      END AS StartSignal
      ,CASE
         WHEN 10 = SUM(CASE WHEN [value] < 1000 THEN 1 ELSE 0 END)
                  OVER (ORDER BY id ROWS BETWEEN CURRENT ROW AND 9 FOLLOWING)
         THEN 1
         ELSE 0
      END AS StopSignal
   FROM patternDemo
)
,PKOnOffs
AS
(
   SELECT id, [value]
      ,CASE
         WHEN StartSignal = 1
         THEN RIGHT('000000000' + CAST(id AS varchar(20)), 10) + 'On'
         WHEN StopSignal = 1
         THEN RIGHT('000000000' + CAST(id AS varchar(20)), 10) + 'Off'
      END AS PKOnOff
   FROM Rules
)
,OnOffs
AS
(
   SELECT id, [value]
      -- MAX ignores NULLs
      ,ISNULL(SUBSTRING(MAX(PKOnOff) OVER (ORDER BY id), 11, 3), 'Off') AS OnOff
   FROM PKOnOffs
)
SELECT id, [value]
FROM OnOffs
WHERE OnOff = 'On';

tricky query with patterns of consecutive rows

Cookies on SQLServerCentral